Hong Kong police: Face scans start, no word on wrong matches

Hong Kong police have solved more than two thousand cases with their camera network. Now they're adding face scanning to it. And so far, nobody has said what happens when the system picks the wrong person.

If you've ever walked through a busy street or a

If you've ever walked through a busy street or a shopping mall, this story is about you. Cameras already watch those places. What's changing is that software will now try to name the faces in them.

It's a little unsettling, and that reaction makes sense.

Hong Kong police say they're switching on A.I. facial recognition across their SmartView camera network this month. They say they'll start with more than three thousand cameras by the end of the year. The plan then grows to about sixty thousand cameras by twenty twenty-eight. So when the software gets it wrong, who finds out, and who fixes it?

Start with what police say the system has done

Start with what police say the system has done. According to the police, SmartView has already helped solve two thousand, one hundred and ninety-two criminal cases. It helped lead to more than two thousand eight hundred arrests. The network today includes about ten thousand police cameras. It also connects to another ten thousand run by other government departments.

That track record is why expansion has support. Cutting a review job from hours to seconds is a real benefit. Police say officers ranked inspector or above will have to sign off on matches.

But a signature only checks the answer the machine hands over. And that's where the research gets uncomfortable.

Studies of these systems show that false matches

Studies of these systems show that false matches can be up to a hundred times more common for Black and Asian people than for white men. Real-world errors can also climb as high as one in five when the footage is poor quality. Street cameras are often exactly that, with bad angles, bad light, and people in motion.

Now add scale. Even a tiny error rate, say one in a thousand, adds up across sixty thousand cameras. Every one of those errors is a real person who looks like someone police want to find.

We already know what that looks like. The Innocence Project reports at least seven confirmed cases of people wrongly identified by facial recognition. Six of them were Black. Seven people, in systems far smaller than what Hong Kong is planning.

There's also a human problem on top of the

There's also a human problem on top of the technical one. When software shows a high confidence score, officers tend to trust it. Some have even called unverified matches a "hundred percent match." That label can steer an entire investigation before anyone checks the facts.

For detectives, that's a process risk. For everyone else, it means a stranger's look-alike face could put you at the center of someone else's case.

The supporters do have a fair argument. The best algorithms on the N.I.S.T. leaderboard, from the U.S. government's testing lab, show very low error rates. Their false matches are held to about three in a thousand. The leading ones also perform well across non-white faces. And studies find that human observers make far more mistakes at the same task.

The question becomes which system Hong Kong

So the question becomes which system Hong Kong actually buys, and how strict its quality rules are. Police haven't answered that in public.

The real issue isn't whether facial recognition works. It's what happens to the people it gets wrong. A system can be mostly accurate and still fail badly if nobody has a clear way to challenge a false match.

Hong Kong is adding face scanning to a huge camera network, and police say it helps solve crimes. The software isn't perfect, and its mistakes fall harder on some groups than others. What's missing is a public answer on how an innocent person would clear their name.